IEEE RELIABILITY SOCIETY SECTION

Received 22 August 2022, accepted 6 September 2022, date of publication 15 September 2022,
date of current version 23 September 2022.
Digital Object Identifier 10.1109/ACCESS.2022.3206954

An Intelligent IoT Sensing System for Rail Vehicle

Running States Based on TinyML
SHAOZE ZHOU , YONGKANG DU, BINGZHI CHEN, YONGHUA LI, AND XINGSEN LUAN
School of Rolling Stock Engineering, Dalian Jiaotong University, Dalian 116028, China
Corresponding author: Shaoze Zhou ([email protected])
This work was supported in part by the China National Natural Science Foundation under Contract 51875073, and in part by the Dalian
Science and Technology Innovation Fund Plan Project under Contract 2019J11CY017.

1 ABSTRACT Real-time identification of the running state is one of the key technologies for a smart rail
2 vehicle. However, it is a challenge to accurately real-time sense the complex running states of the rail
3 vehicle on an Internet-of-Things (IoT) edge device. Traditional systems usually upload a large amount of
4 real-time data from the vehicle to the cloud for identification, which is laborious and inefficient. In this paper,
5 an intelligent identification method for rail vehicle running state is proposed based on Tiny Machine Learning
6 (TinyML) technology, and an IoT system is developed with small size and low energy consumption. The
7 system uses a Micro-Electro-Mechanical System (MEMS) sensor to collect acceleration data for machine
8 learning training. A neural network model for recognizing the running state of rail vehicles is built and trained
9 by defining a machine learning running state classification model. The trained recognition model is deployed
10 to the IoT edge device at the vehicle side, and an offset time window method is utilized for real-time state
11 sensing. In addition, the sensing results are uploaded to the IoT server for visualization. The experiments
12 on the subway vehicle showed that the system could identify six complex running states in real-time with
13 over 99% accuracy using only one IoT microcontroller. The model with three axes converges faster than the
14 model with one. The model recognition accuracy remained above 98% and 95%, under different installation
15 positions on the rail vehicle and the zero-drift phenomenon of the MEMS acceleration sensor, respectively.
16 The presented method and system can also be extended to edge-aware applications of equipment such as
17 automobiles and ships.

18 INDEX TERMS TinyML, IoT, running state, smart rail vehicle, artificial neural network.

19 I. INTRODUCTION achieved by the real-time acquisition of vehicle-side ECU 32

20 Enabling rail vehicles to have self-awareness through sensors sensors or GPS data which is sent to the cloud when the 33

21 with low energy consumption is a challenge as the rail vehicle vehicle is moving. States results are fed back to the vehicle 34

22 industry is rapidly developing towards intelligence and low after using complex algorithms in the cloud to identify the 35

23 carbonization. Real-time identification of the running state is various states of the vehicle. This type of system is not suit- 36

24 the key technology for realizing the self-awareness of smart able for smart rail vehicles because of its disadvantages and 37

25 rail vehicles. However, identifying various complex running limitations such as high cost, high power consumption, large 38

26 states with low energy consumption is a difficult task because size, poor real-time performance, and complicated structure. 39

27 the computing power of edge devices at the vehicle side is low In contrast, the edge state sensing system with small size, 40

28 and the state data monitored by sensors is complex. low cost, and low power consumption has high practical and 41

29 Currently, state monitoring is mostly focused on the auto- economic value for intelligent rail vehicles. For example, the 42

30 motive domain in the existing studies [1], [2], [3], [4], [5], system can record the running states of each vehicle, and 43

31 [6]. Most state monitoring systems of these researches are provide accurate and quantitative data for structural health 44

monitoring and condition-based maintenance of rail vehi- 45

The associate editor coordinating the review of this manuscript and cles. In addition, this type of system can provide long-term 46

approving it for publication was Zhaojun Steven Li . tracking and monitoring services for rail vehicles (such as 47

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

98860 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 10, 2022
 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 1. Schematic diagram of the Intelligent IoT sensing system for rail vehicle running states based on TinyML.

48 rail wagons) only using solar power. Therefore, carrying rail vehicles has not been reported. In addition, the research 89

49 out the research on edge identification technology will have on which sensor data (such as speed, acceleration, or gyro- 90

50 important theoretical research and engineering application scope data) can accurately determine the running state of rail 91

51 significance to the design, manufacture, and maintenance of vehicles should be conducted. Since the state data recorded 92

52 smart rail vehicles. by rail vehicle operation has its uniqueness, it is essential to 93

53 With the development of artificial intelligence, some analyze its characteristics, and study state recognition meth- 94

54 researchers have used machine learning in automobile vehicle ods and machine learning models. 95

55 intelligence [7], [8]. These studies include autonomous vehi- This paper proposes a novel intelligent monitoring and 96

56 cle driving [9], [10], intelligent vehicle classification [11], identification method for rail vehicles’ running states based 97

57 and intelligent vehicle monitoring [12]. Some researchers on TinyML technology. An IoT intelligent sensing system 98

58 [13], [14] considered vehicle running safety on vehicle intelli- was developed with a small size, low cost, and low energy 99

59 gent classification, which classified and identified the vehicle consumption. Only a miniature IoT edge device was used 100

60 running states. Some researchers have studied the state recog- to identify multiple complex running states. Taking subway 101

61 nition methods for various equipment. Lan et al. [15] repre- vehicle monitoring as an example, the system realized six 102

62 sented a diagnosis strategy based on operating conditions and kinds of real-time running state identification effectively and 103

63 pressure pulsation of the turbine in order to effectively mon- transmitted the results to the IoT server for visualization. 104

64 itor the operating state of hydraulic turbines. Lu et al. [16] In order to evaluate the system’s reliability and effectiveness, 105

65 proposed a GA-CNN model to achieve automatic recognition experiments were also conducted with different numbers of 106

66 of the rolling bearing running state. acceleration axes, different deployment positions, and zero- 107

67 A number of studies already had been conducted to apply drift effects. 108

68 machine learning in combination with microcontrollers.

69 Adhitya et al. [17] used microcontrollers that read two inputs II. SYSTEM ARCHITECTURE 109

70 from temperature and humidity sensor data and outputted The Intelligent IoT sensing system for rail vehicle running 110

71 two neurons to control two actuators by a machine learning states based on TinyML consists of a data acquisition module, 111

72 model algorithm. Chand et al. [18] designed a system based a machine learning module, a real-time state recognition 112

73 on monitoring data of ultrasonic sensors and used machine module, and a server module. The system architecture dia- 113

74 learning algorithms to analyze the data and analyze personnel gram is shown in Figure 1. In this paper, acceleration signals 114

75 behavior. Zhang et al. [19] investigated the deployment of are tried to identify and classify the running state of vehicles. 115

76 convolutional neural networks on an embedded platform to

77 implement a target detector. In recent years, edge machine A. THE DATA ACQUISITION MODULE 116

78 learning has developed rapidly and Tiny Machine Learning The data acquisition module mainly collects acceleration data 117

79 (TinyML) technology becomes an increasing area [20], [21], for machine learning training. Figure 2 illustrates the mod- 118

80 [22]. Prado et al. [23] proposed a TinyML-based method ule hardware consists of an ESP32 microcontroller, a low- 119

81 to apply a machine vision algorithm to mini-car automatic cost Micro-Electro-Mechanical System (MEMS) accelera- 120

82 driving. tion ADXL345 sensor, a clock module, an SD card memory 121

83 So far, the above research literature has studied the creation module, a battery, and a circuit board. The module size is only 122

84 or comparative optimization of automotive and device- 90mm × 28 mm × 38 mm and the power consumption is only 123

85 specific machine learning models. However, in the trail vehi- 7.26E-3W. When the system is initialized, the ESP32 system 124

86 cle field, research on Internet-of-Things (IoT) edge machine time is updated from the Internet or local area network NTP 125

87 learning models is still lacking. In particular, research on server. The acceleration value and time stamp are recorded on 126

88 intelligent state recognition based on TinyML technology for the SD card in the module. 127

VOLUME 10, 2022 98861

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 2. The hardware of the data acquisition or the real-time running

state recognition module.

128 B. THE MACHINE LEARNING MODULE

129 The learning module is constructed by an open-source frame-
130 work such as Tensorflow. The machine learning framework
131 is built with the artificial neural network, and the collected
132 acceleration data is used for training. After training, a model
133 file to identify the running state features is obtained. The
FIGURE 3. The workflow chart of the Intelligent IoT sensing system for
134 model is converted into TinyML model format and deployed rail vehicle running states based on TinyML.
135 to the IoT microcontroller.

136 C. THE REAL-TIME RUNNING STATE RECOGNITION the collected data is segmented, normalized, dimensionally 161
137 MODULE transformed, and labeled with defined values. 162

138 The real-time state monitoring module has the same hardware Step 3: Build a machine learning artificial neural net- 163

139 configuration as the data acquisition module. But the internal work, and generate a model for identifying the running 164

140 software and functions are different. After deploying the state features. The running state sample set is expanded 165

141 TinyML model file to the IoT microcontroller on the vehicle into one-dimensional data and imported into the input layer. 166

142 side, the acceleration values are read from the sensor and the Furthermore, the artificial neural network is constructed by 167

143 vehicle’s running state is identified by the model in real-time. setting proper parameters such as loss function, activation 168

function, and training rounds. After training, a model is 169

144 D. THE SERVER MODULE generated to identify the running state features. The model 170

145 Using the programmability of the edge device ESP32 micro- prediction accuracy is used as the basis for assessing the 171

146 controller chip, recognition results are sent to the local or model quality. 172

147 remote cloud IoT servers, and other data servers through the Step 4: Identify the running state in real time. the identi- 173

148 WebSocket protocol. The real-time results are accessed by fication of the running state feature model is converted and 174

149 browsers for terminal users. deployed to the IoT microcontroller based on TinyML tech- 175

nology. The microcontroller monitors the acceleration in real 176

150 III. METHODOLOGICAL STEPS time as inputs to the TinyML model for inference recognition. 177

151 The overall workflow diagram of the Intelligent IoT sensing The recognition results are sent by the microcontroller to the 178

152 system for rail vehicle running states based on TinyML is IoT server. 179

153 displayed in Figure 3. The specific steps of the proposed The key technologies of the above steps are addressed in 180

154 method are as follows. detail below. 181

155 Step 1: Collect the acceleration data of the complete start-

156 stop cycles. The data acquisition module collects sufficient X, A. PREPROCESSING OF RUNNING STATE ACCELERATION 182

157 Y, and Z-axis acceleration data of complete start-stop cycles DATA 183

158 from the rail vehicle. Step 2 is the preprocessing of acceleration data, including 184

159 Step 2: Preprocess the data that has been collected. To gen- data segmentation, normalization, dimensional transforma- 185

160 erate various running state sample sets for machine training, tion, and label definition. 186

98862 VOLUME 10, 2022

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 4. The typical one start-stop cycle of the rail vehicle.

FIGURE 5. The example of the training set.

187 1) DATA SEGMENTATION

188 Figure 4 shows the acceleration values of a typical rail vehicle objects are fused [24], [25], [26]. To avoid this situation, Eq. 227

189 in one complete start-stop cycle. The X-axis is the longitu- (1) normalization is used in machine learning tasks to ensure 228

190 dinal direction of the vehicle, while the Y-axis and Z-axis that each feature magnitude has an equal contribution when 229

191 are the lateral direction and vertical direction, respectively. fed to the classifier and to eliminate the effect of odd data. 230

192 The vehicle moved along the X-axis direction. The figure After normalization, the data is distributed between 0 and 1, 231

193 presents that the X-axis acceleration values of the rail vehicle and the optimization of the data is faster and the accuracy is 232

194 are positive when accelerating, negative when decelerating, higher during training. 233

195 and steady and close to zero when stationary. According xi − xmin
196 to the acceleration values, the running states of some key xi0 = (1) 234
xmax − xmin
197 positions in the figure are presented as follows: 1) starting,
198 2) accelerating, 3) moving, 4) decelerating, 5) stopping, and where xi0 is normalized data, while xi is collected data, and 235

199 6) stationary. These running state accelerations have distinct xmax and xmin represent the maximum value and the minimum 236

200 characteristics. Taking the X-axis acceleration as an example, value of the M rows of data, respectively. 237

201 the acceleration values of the starting state is from small to

202 large. The acceleration values of the moving state have a wide 3) DIMENSIONAL TRANSFORMATION 238

203 range of fluctuation around the zero axes. In the decelerating To simplify the model training process, the three axes data 239

204 state, the acceleration values are negative. of the Sn group data are connected end to end to form a one- 240

205 The acceleration data is divided into sample sections of dimensional neural array L of 3 × M values. This array is 241

206 different running state types, and each sample section has used as the input of the neural network. 242

207 the same time interval Tg , such as 2 seconds. Each sample

208 contains three columns and M rows of data for X, Y, and 4) LABEL VALUE DEFINITION 243

209 Z-axis accelerations. As shown in Figure 4, section A is Since the machine learning model is based on a multi- 244

210 the accelerating type sample, which has obvious acceleration classification learning algorithm, one-hot code [27] is 245

211 characteristics. The same types of samples are segmented, adopted to identify classification information for different 246

212 extracted, and combined to form a sample set of Z rows, running state data. For example, to identify the three running 247

213 as in Figure 5. Then Z equals M × N , where N is the states of rail vehicles in accelerating, decelerating, and mov- 248

214 number of samples. Sn (n = 1,2,3. . .) is defined as the name of ing, the three labels are defined as 0, 1, and 2. The one-hot 249

215 each sample. Add the Z rows sample set with the state label encoding forms are [1,0,0], [0,1,0], and [0,0,1]. The reason 250

216 name Tn (n = 1,2,3. . .). For example, T1 is the accelerating for using one-hot coding is that it extends the value of discrete 251

217 sample set and T2 is the decelerating sample set. To ensure features to Euclidean space. 252

218 the reliability of the machine learning model and to determine

219 whether the training degree is overfitting or underfitting, the B. TINYML RUNNING STATE RECOGNITION MODEL 253

220 neural network dataset is divided into a training set and a TRAINING 254

221 verification set. The ratio of the training set to the validation In Step 3, running state recognition is implemented using 255

222 set is 4:1. a classification machine algorithm, which is modeled and 256

trained using the sequence model provided by the Keras 257

223 2) NORMALIZATION library of artificial intelligence framework, as displayed in 258

224 In the classification algorithm, the classification performance Figure 6. Relevant parameters such as loss function, acti- 259

225 of the learning algorithm will be reduced when the attributes vation function, and optimization algorithm are adjusted 260

226 and different feature information of multiple description according to the model training condition. The sample set 261

VOLUME 10, 2022 98863

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 6. The neural network structure of the running state recognition

model.

FIGURE 7. Schematic diagram of vehicle running state data time offset

262 processed in subsection A is utilized as the input of the neural window.
263 network. If the model needs to recognize K types of running
264 states, the number of output layer units is set to K .
265 The vehicle running state model belongs to the multi- is completed, the optimal values are selected as the final 299

266 classification machine learning algorithm. The Softmax func- parameter values. The number of neurons in the hidden layer 300

267 tion is chosen for the activation function from the hidden layer is decided by Eq. (4). 301

268 to the output layer of this model. Its expression is presented p

N = Ni (No + 1) + 1 (4) 302
269 in Eq. (2).
where N is the number of neurons in the hidden layer, and 303
exp(xi )
270 softmax (xi ) = P (2) Ni and No represent the number of neurons in the input 304

j exp(xj ) and output layers, respectively. If the training result is not 305

271 The numerator in equation (4) is the exponential function satisfactory, adjust the model parameters and retrain. 306

272 of the signal xi received by the neurons in the output layer,

C. PERFORMANCE EVALUATION METHOD FOR RUNNING 307
273 and the denominator sums the exponential functions of all the
STATE RECOGNITION MODELS 308
274 output results xj . Each output result of the neural network is
275 presented in the form of a probability distribution. In the training task in Step 3, the machine learning mod- 309

276 In a multi-classification problem based on the Softmax els obtained differ due to the selected parameters. Ideally, 310

277 classifier, the loss function is used to define the effect of the generalization error of all candidate models should be 311

278 the neural network model and the goal of optimization. The compared to find the optimal model. However, since the 312

279 cross-entropy loss function is adopted in this study. The for- generalization error cannot be obtained directly, this paper 313

280 mula for calculating the cross entropy from the sample output evaluates the model by its accuracy. 314

281 of the machine learning model with the actual label values is For the classification task, the sample error rate and accu- 315

282 shown in Eq. (3). racy rate are defined as (5) and (6). 316

1 Xm Y
E(f ; D) = (f (xi ) 6= yi ) (5)
X
283 Hy0 (y) = − y0i log(yi ) (3) m i=1
317

i 1 Xm Y
acc(f ; D) = (f (xi ) = yi ) = 1 − E(f ; D) (6) 318
284 where yi is the probability of each result output by the model, m i=1
285 and y0i represents the category corresponding to the one-hot where D is the sample set and m represents the total number 319

286 code. of samples. In addition to the accuracy, the loss function also 320

287 The Adam optimization algorithm, which combines the is used to evaluate the recognition model. 321

288 features of the Adagrad and RMSprop algorithms, is adopted

289 in the machine learning of the system. D. DEPLOYMENT AND REAL-TIME RECOGNITION OF 322

290 The labeled samples from subsection A are used for TINYML MODEL 323

291 machine learning to determine the ideal values of all weights After completing the recognition model training in Step 4, 324

292 and deviations. When constructing the model, the parameters the machine learning model file is compressed as a TinyML 325

293 of machine learning need to be determined. Among them, the model and deployed to the MCU of the IoT edge device in 326

294 numbers of neurons in the input layer and the output layer the vehicle. The edge device sensor monitors the running 327

295 are determined according to the number of identified running accelerations and transmits them to the TinyML model for 328

296 states. Other parameters to be determined are the number of real-time discrimination. 329

297 hidden layers, the number of neurons in the hidden layer, the To prevent the state information from being lost because 330

298 number of training rounds, etc. When the model evaluation the intercepted signal misses the characteristic signal, 331

98864 VOLUME 10, 2022

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 8. The subway vehicle experiment for testing the Intelligent IoT
sensing system.

332 this paper presented an offset time window method. Fig-

333 ure 7 shows a start-stop cycle of the rail vehicle. When in real-
334 time monitoring, the data is segmented and moved forward
335 with Tf as the time offset window and Tg as the recognition
336 time interval for each acquisition. For example, Tf takes
337 0.5 seconds and Tg takes 2 seconds, which means that every
338 0.5 seconds, 2 seconds of acceleration data is intercepted for
339 identification. As a result, the offset time window method
340 will repeatedly recognize the running state. For statistical
341 convenience, the results of these consecutive repetitions are
342 combined into one. The process described above is an online
343 real-time identification method, which is also applicable to FIGURE 9. Acceleration sample data of two start-stop cycles (a) X-axis
acceleration data (b) Y and Z-axis acceleration data.
344 offline recognition of recorded acceleration data. Through the
345 offline method, the accuracy of the recognition model can
346 be tested to evaluate the quality of the trained model. The
347 recognition result and time stamp will be recorded.

348 IV. EXPERIMENTS

349 The effectiveness of the constructed model and system was
350 verified by a real-time recognition experiment for subway
351 running states. The machine learning framework adopted the
352 open-source framework TensorFlow. As presented in Fig-
353 ure 8, an acquisition module was placed in the front part of
354 a subway carriage to acquire acceleration values for training.
355 After training, a state recognition module was used for real-
356 time recognition.
FIGURE 10. Schematic diagram of X-axis acceleration of three running
357 A. ACCELERATION DATA ACQUISITION states.
358 A data acquisition module with an acquisition frequency of
359 20Hz was used to collect acceleration data from the subway
360 vehicle. To ensure sufficient data, the test collected more than were set as T1 , T2 , and T3 (K equals 3) as training data. 372

361 100,000 rows of three-axis acceleration data. In Figure 9, the Figure 10 is the schematic diagram of the X-axis acceleration 373

362 acceleration data of two start-stop cycles at the head of the of three states for one start-stop cycle without the stationary 374

363 subway carriage. The subway traveled along the longitudinal state. Using M equals 40 and Z equals 100, the collected 375

364 X-axis direction of the accelerometer. Figure 9a) shows a sig- three-axis acceleration data were preprocessed according to 376

365 nificant change in acceleration in the direction. The vehicle the proposed method. Figure 11 illustrates some collected 377

366 in Figure 9b) also reveals vibration in the Y and Z axes with data. 378

367 a certain pattern during running. Because the stationary state has a single characteristic and 379

is easily judged programmatically, the three running states 380

368 B. TRAINING DATA PREPROCESSING do not contain the stationary state. The specific judgment 381

369 To identify three running states, the data sets of three running method of the stationary state is that the three-axis acceler- 382

370 states of rail vehicles in accelerating, moving, and deceler- ation values are always less than a small threshold range in 383

371 ating were extracted from the collected data. The data sets the Tg section. 384

VOLUME 10, 2022 98865

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 12. Schematic diagram of five different vehicle running states

recognition.

FIGURE 11. Part of the sample set of subway (a) sample set of the FIGURE 13. Training and recognition results of five vehicle running states
accelerating state (b)sample set of the decelerating state. by the model.

TABLE 1. The machine learning model’s parameters for recognizing three

running states. the evaluation accuracy of the training set was 100%. The 396

machine learning model was used for offline prediction with 397

4640 rows of acceleration data of the vehicle running for two 398

complete cycles, and the prediction accuracy of the offset 399

time window method was 99.8%. 400

Furthermore, the proposed method was verified with five 401

complex types: (1) starting, (2) accelerating, (3) moving, (4) 402

decelerating, and (5) stopping. Figure 12 presents the X-axis 403

acceleration data for the five running states. The parameters 404

of the machine learning model for five running states are 405

presented in Table 2. After the model training, the accuracy 406

of the training set evaluation is 100%, and the accuracy of the 407

validation set is 99.67%, as shown in Figure 13. 408

385 Based on the method in subsection III, the segmentation The model was used to predict the offline data of three 409
386 data was fed into the neural network as input neurons. The complete cycles. A total of 6520 acceleration data points were 410
387 three-dimensional array of triaxial acceleration Sn with M predicted 652 times using the offset time window method. 411
388 equals 40 was flattened into a 120 × 1 one-dimensional array The results had seven errors, and the model accuracy was 412
389 as the input to the neural network. After defining the three 98.9%. Most of the errors were identifying acceleration state 413
390 running states as 0, 1, and 2 labels, the data was normalized. features as starting states. The reason for this is that some 414

acceleration data had very similar characteristics of acceler- 415

391 C. COMPLEX RUNNING STATE RECOGNITION AND ating and starting in the Tg time interval. This problem can 416

392 TRAINING be eliminated by determining whether the previous state of 417

393 The preprocessed sample data set was trained for machine the current state is stationary or not. If the previous Tg state 418

394 learning. The parameters of the machine learning model for is stationary, the current running state is starting, and if it is 419

395 three running states are listed in Table 1. After training, not, it is accelerating. 420

98866 VOLUME 10, 2022

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 14. Comparison of accuracy of different activation functions (a)M=5 (b)M=10 (c)M=20 (d)M=40.

TABLE 2. The machine learning model’s parameters for recognizing five TABLE 3. Statistics on the training effect of different activation function
running states. models.

tion. When the number of input rows gradually increases (M 436

equals 40), the model training effect of the Relu function and 437
421 D. THE TRAINING EFFECT OF VARIOUS ACTIVATION
the Tanh function has been greatly reduced. When using the 438
422 FUNCTION MODELS
Sigmoid activation function, the accuracy of the model does 439
423 The activation function in a neural network can improve the not decrease with the number of input M . The model is robust 440
424 nonlinear fitting ability of the model, which is crucial to the when using the Sigmoid function as the activation function, 441
425 quality of the trained model. To choose a suitable activation and the system has a more stable recognition capability. The 442
426 function, three common activation functions were compared number of training rounds changes as the number of input 443
427 by giving different M rows of data. Three common activation rows increases, as shown in Table 3. 444
428 functions were applied to the model for 50 rounds of train-
429 ing, and the accuracy of the training results is presented in
430 Figure 14. E. TRAINING EFFECT COMPARISON ON DIFFERENT AXIS 445

431 Figure 14 reveals that the training results using different NUMBERS 446

432 activation functions are different when M is used as the vari- Since the vehicle running state features in Figure 9a) are 447

433 able. When M equals 5 and 10, all three activation functions already visible on the X-axis, the effect of comparing one 448

434 achieve good learning results. The model training speed is X-axis and three-axis acceleration data as input to the model 449

435 faster when using the Relu function as the activation func- recognition must be investigated. 450

VOLUME 10, 2022 98867

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 15. Comparison of the training effects of different axes with

M=20. FIGURE 16. Comparison of the training effects of different axes with
M=20.

451 Acceleration data with two different numbers of axes were

452 used separately as inputs for model training. When M equals
453 20, the resultant accuracy and cross-entropy loss values are
454 described in Figure 15.
455 The same activation function, optimization function, and
456 other parameters were chosen for the construction of the two
457 machine learning models. The model in figure 15 has been
458 trained after 20 iterations when the three-axis acceleration
459 data is taken as the model input. In contrast, using the X-axis
460 acceleration data as input takes more than 50 training rounds.
461 Consequently, the model with three axes converged faster
462 with higher accuracy and lower loss value than the model with FIGURE 17. Comparison of the accuracy of different monitoring positions.
463 one axis. Therefore, the system chose three-axis acceleration
464 as the input to train the model of the running state.

465 F. COMPARISON OF RECOGNITION ACCURACY FOR

466 DIFFERENT INSTALLATION POSITION
467 The length of a subway carriage is generally long. A typical
468 subway carriage is 19 to 23 meters. To test the reliability,
469 robustness, and generalization ability of the system, the accu-
470 racy influence of different installation positions of the system
471 in the vehicle needs to be studied. In the experiments, the
472 recognition model was trained with acceleration data from
473 the head of the carriage only and without training by the data
474 from the rear of the carriage. A sample of the collected data
475 from the rear of the subway carriage is shown in Figure 16. FIGURE 18. The acceleration data with zero drift.
476 The figure reveals that the rear acceleration data have the
477 same characteristics as the head one in Figure 9.
478 Experiments were conducted to identify the three running module can be deployed in different positions in the same 489

479 states of accelerating, moving, and decelerating as an exam- carriage. 490

480 ple. After splitting and defining the label value, the dataset
481 with Z equals 1000 was imported into the model for recogni- G. COMPARISON OF THE ACCURACY OF ZERO-DRIFT 491

482 tion evaluation. The accuracy rate of the evaluation results is RECOGNITION 492

483 demonstrated in Figure 17. Since zero drift may occur in low-cost MEMS sensors, it is 493

484 After model identification, the three states were identified necessary to analyze the effect of this phenomenon on recog- 494

485 with 99.8%, 98.2%, and 100% accuracy. The high recogni- nition results if zero drift is not removed. 495

486 tion accuracy of the model demonstrated that the installation As shown in Figure 18, the data with z-axis zero drift (Z 496

487 position of the state recognition module in the carriage had equals 1000) was imported into the model for offline identi- 497

488 little effect on the accuracy. Therefore, the state recognition fication. The recognition results are presented in Figure 19. 498

98868 VOLUME 10, 2022

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

FIGURE 19. Comparison of model evaluation with zero-drift acceleration

data. FIGURE 20. The web interface of real-time recognition results of running
states.

499 The recognition accuracies of accelerating, decelerating, and

500 moving states are 99.5%, 99%, and 95.1%. Therefore, the there were three abnormalities in 1108 times of recognition 538

501 recognition model has strong generalization ability and reli- result data, with an accuracy rate of 99.7%. 539

502 ability because it can identify the acceleration data with zero
503 drift to a certain extent. On the other hand, it also reveals V. CONCLUSION 540

504 that the zero-drift decreased the identification accuracy and TinyML technology is an effective way to make rail 541

505 had a certain influence on the recognition results. Thus, the vehicles intelligent and perceptive. The constructed sys- 542

506 zero-drift needs to be removed for obtaining better recogni- tem is low-cost, small-sized, and low energy consump- 543

507 tion results. tion. With only one micro IoT edge device on the vehicle 544

side, each carriage can recognize complex running states. 545

508 H. REAL-TIME SENSING OF RUNNING STATES FOR THE The following conclusions are obtained from this paper’s 546

509 RAIL VEHICLE ONLINE research: 547

510 The TinyML recognition model was deployed to the IoT (1) The recognition model and system proposed in this 548

511 edge device on the vehicle to sense the vehicle running paper can identify multiple complex running states in real- 549

512 states online. Then, the effectiveness of the system was tested time, and the monitoring results are accurate. The recognition 550

513 by sensing six types of complex motion states: starting, model is robust when using the Sigmoid function as the 551

514 accelerating, decelerating, stopping, moving, and stationary. activation function. The subway experiments showed that the 552

515 Among the six states, the stationary state was identified by the system was capable of identify six complex running states 553

516 method in subsection IV(B), while the rest of the states were in real-time and the identification accuracy was higher than 554

517 identified by the machine learning method. Four start-stop 99%. Moreover, the experiment shows that the acceleration 555

518 cycles of the subway were monitored in the experiment. The data can be used for monitoring data for identifying the 556

519 total real-time monitoring time was about 9 minutes. Each running state of smart rail vehicles. 557

520 start-stop cycle had a deceleration time of about 1 minute, (2) Three-axis acceleration data monitoring is better than 558

521 a stop time of about 1.5 minutes, and an acceleration time of one-axis monitoring. Using three-axis acceleration data as 559

522 about 20 seconds. model input for model training is faster, more accurate, 560

523 The system used the online offset time window method and has lower loss values than using one-axis acceleration 561

524 for intelligent identification. To verify the accuracy of the data. 562

525 recognition results, besides online intelligent recognition, (3) The sensing system can obtain high accuracy in run- 563

526 the whole test process was also recorded on video for ning state recognition regardless of whether it is installed 564

527 offline rechecking. The Tf of the experiment was taken as in the front or rear of the carriage. The experiment on 565

528 0.5 seconds and Tg was taken as 2 seconds, which means the subway revealed that the recognition accuracy of dif- 566

529 that running states were recognized every 0.5 seconds and ferent installation positions was high, reaching more than 567

530 were judged by 2 seconds of acceleration data. A total of 98%. The model has good reliability and generalization 568

531 1108 judgments were made in the whole test. During sensing, capability. 569

532 the recognition results were uploaded to the local IoT server. (4) The established model retains some high-running 570

533 Each complex running state of the vehicle was presented on state recognition capability under zero-drift acceleration. The 571

534 the browser in real time, as shown in Figure 20. After online experimental result showed that the recognition accuracy of 572

535 and offline result comparison, the test results revealed that the model was still high, reaching more than 95%, with- 573

536 the real-time recognition results using the proposed method out removing zero drift. Furthermore, it indicated that the 574

537 matched with the subway vehicle running status results, and zero-drift has a certain influence on the recognition accuracy 575

VOLUME 10, 2022 98869

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

576 and the zero-drift should be removed during monitoring for [15] C. Lan, S. Li, H. Chen, W. Zhang, and H. Li, ‘‘Research on running state 642

577 obtaining better results. recognition method of hydro-turbine based on FOA-PNN,’’ Measurement, 643
vol. 169, Feb. 2021, Art. no. 108498. 644
578 The method proposed in this paper can identify more [16] W. Lu, H. Mao, F. Lin, Z. Chen, H. Fu, and Y. Xu, ‘‘Recognition 645
579 complex running states such as turning, collision, rapid accel- of rolling bearing running state based on genetic algorithm and con- 646

580 eration, slow acceleration, emergency braking, etc. In addi- volutional neural network,’’ Adv. Mech. Eng., vol. 14, no. 4, 2022, 647
Art. no. 168781322210956. 648
581 tion, the system is suitable for mass installation on rail [17] R. Y. Adhitya, M. A. Ramadhan, S. Kautsar, N. Rinanto, S. T. Sarena, 649
582 vehicles. The identification results can provide essential I. Munadhif, M. Syai’in, R. T. Soelistijono, and A. Soeprijanto, ‘‘Com- 650

583 data support for structural health monitoring and condition- parison methods of fuzzy logic control and feed forward neural network 651
in automatic operating temperature and humidity control system (oyster 652
584 based maintenance. The method and system can also be mushroom farm house) using microcontroller,’’ in Proc. Int. Symp. Elec- 653
585 applied to various rail vehicles like high-speed trains, rail- tron. Smart Devices (ISESD). Piscataway, NJ, USA: IEEE, Nov. 2016, 654

586 way wagons, and intercity trains, as well as other equip- pp. 168–173. 655
[18] G. Chand, M. Ali, B. Barmada, V. Liesaputra, and G. Ramirez-Prado, 656
587 ment with TinyML chips such as cars, ships, aircraft, and ‘‘Tracking a person’s behaviour in a smart house,’’ in Proc. 657
588 spacecraft. Service-Oriented Comput. (ICSOC), in Lecture Notes in Computer 658
Science, X. Liu, Eds. Cham, Switzerland: Springer, 2019, 659
pp. 241–252. 660
589 REFERENCES [19] Y. Zhang, S. Bi, M. Dong, and Y. Liu, ‘‘The implementation of CNN-based 661
object detector on ARM embedded platforms,’’ in Proc. IEEE 16th Int. 662
590 [1] S. J. Kamble and M. R. Kounte, ‘‘Machine learning approach on traffic
Conf. Dependable, Autonomic Secure Comput., 16th Int. Conf. Pervasive 663
591 congestion monitoring system in internet of vehicles,’’ Proc. Comput. Sci.,
Intell. Comput., 4th Int. Conf. Big Data Intell. Comput. Cyber Sci. Technol. 664
592 vol. 171, pp. 2235–2241, Jan. 2020.
Congr. (DASC/PiCom/DataCom/CyberSciTech), Athens, Greece, 82018, 665
593 [2] J. G. Choi, C. W. Kong, G. Kim, and S. Lim, ‘‘Car crash detection using pp. 379–382. 666
594 ensemble deep learning and multimodal data from dashboard cameras,’’ [20] E. Flamand, D. Rossi, F. Conti, I. Loi, A. Pullini, F. Rotenberg, and 667
595 Expert Syst. Appl., vol. 183, Nov. 2021, Art. no. 115400. L. Benini, ‘‘GAP-8: A RISC-V SoC for AI at the edge of the IoT,’’ in 668
596 [3] T. Praveenkumar, B. Sabhrish, M. Saimurugan, and K. I. Ramachandran, Proc. IEEE 29th Int. Conf. Appl.-Specific Syst., Archit. Processors (ASAP), 669
597 ‘‘Pattern recognition based on-line vibration monitoring system for fault Jul. 2018, pp. 1–4. 670
598 diagnosis of automobile gearbox,’’ Measurement, vol. 114, pp. 233–242, [21] P. Warden and D. Situnayake, TinyML. Sebastopol, CA, USA: O’Reilly 671
599 Aug. 2018. Media, 2019. 672
600 [4] Q. Lu, L. Huang, W. Li, T. Wang, H. Yu, X. Hao, X. Liang, F. Liu, [22] I. Fedorov, R. P. Adams, M. Mattina, and P. Whatmough, ‘‘SpArSe: 673
601 P. Sun, and G. Lu, ‘‘Mixed-potential ammonia sensor using Ag dec- Sparse architecture search for CNNs on resource-constrained micro- 674
602 orated FeVO4 sensing electrode for automobile in-situ exhaust envi- controllers,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 32, 2019, 675
603 ronment monitoring,’’ Sens. Actuators B, Chem., vol. 348, Dec. 2021, pp. 1–13. 676
604 Art. no. 130697. [23] M. de Prado, M. Rusci, A. Capotondi, R. Donze, L. Benini, and N. Pazos, 677
605 [5] S. Liu, N. Guan, D. Ji, W. Liu, X. Liu, and W. Yi, ‘‘Leaking your engine ‘‘Robustifying the deployment of tinyML models for autonomous mini- 678
606 speed by spectrum analysis of real-Time scheduling sequences,’’ J. Syst. vehicles,’’ Sensors, vol. 21, no. 4, p. 1339, 2021. 679
607 Archit., vol. 97, pp. 455–466, Aug. 2019. [24] B. Saha and D. Srivastava, ‘‘Data quality: The other face of big data,’’ in 680

608 [6] A. O. V. P. Kumar, D. Nandini, M. M. Sairam, and B. P. Madhusudan, Proc. IEEE 30th Int. Conf. data Eng., Mar. 2014, pp. 1294–1297. 681

609 ‘‘Development of GPS & GSM based advanced system for track- [25] P. J. Rousseeuw and M. Hubert, ‘‘Robust statistics for outlier detec- 682

610 ing vehicle speed violations and accidents,’’ Mater. Today, Proc., tion,’’ Data Mining Knowl. Discovery, vol. 1, no. 1, pp. 73–79, 683

611 Jul. 2021. 2011. 684

612 [7] S. Lee, Y. Kim, H. Kahng, S.-K. Lee, S. Chung, T. Cheong, K. Shin, J. Park, [26] D. Singh and B. Singh, ‘‘Investigating the impact of data normalization 685

613 and S. B. Kim, ‘‘Intelligent traffic control for autonomous vehicle systems on classification performance,’’ Appl. Soft Comput., vol. 97, Dec. 2020, 686

614 based on machine learning,’’ Expert Syst. Appl., vol. 144, Apr. 2020, Art. no. 105524. 687

615 Art. no. 113074. [27] P. Rodríguez, M. A. Bautista, J. Gonzàlez, and S. Escalera, ‘‘Beyond 688
one-hot encoding: Lower dimensional target embedding,’’ Image Vis. Com- 689
616 [8] M. N. Khan and M. M. Ahmed, ‘‘Trajectory-level fog detection based on
put., vol. 75, pp. 21–31, Jul. 2018. 690
617 in-vehicle video camera with TensorFlow deep learning utilizing SHRP2
618 naturalistic driving data,’’ Accident, Anal. Prevention, vol. 142, Jul. 2020,
619 Art. no. 105521.
620 [9] D. A. Alghmgham, G. Latif, J. Alghazo, and L. Alzubaidi, ‘‘Autonomous
621 traffic sign (ATSR) detection and recognition using deep CNN,’’ Proc.
622 Comput. Sci., vol. 163, pp. 266–274, Jan. 2019.
623 [10] Z. Zhu, Z. Hu, W. Dai, H. Chen, and Z. Lv, ‘‘Deep learning for autonomous
624 vehicle and pedestrian interaction safety,’’ Saf. Sci., vol. 145, Jan. 2022,
625 Art. no. 105479.
626 [11] B. R. Vasconcellos, M. Rudek, and M. de Souza, ‘‘A machine
627 learning method for vehicle classification by inductive waveform
628 analysis,’’ IFAC-PapersOnLine, vol. 53, no. 2, pp. 13928–13932,
629 2020.
630 [12] S. Jaiswal, D. K. Sharma, T. Jaiswal, B. Basumatary, M. Tiwari, and
631 T. Tiwari, ‘‘Real time analysis of intelligent placing system for vehi-
632 cles using IoT with deep learning,’’ Materials Today, Proc., vol. 51, SHAOZE ZHOU received the Ph.D. degree in 691

633 pp. 339–343, Jan. 2022. mechanical engineering from Dalian Jiaotong Uni- 692

634 [13] T. Yonezawa, I. Arai, T. Akiyama, and K. Fujikawa, ‘‘Random forest based versity, Dalian, China, in 2011. He is currently 693

635 bus operation states classification using vehicle sensor data,’’ in Proc. IEEE an Associate Professor with the School of Loco- 694

636 Int. Conf. Pervasive Comput. Commun. Workshops (PerCom Workshops), motive and Rolling Stock Engineering, Dalian 695

637 Mar. 2018, pp. 747–752. Jiaotong University. He has published more than 696

638 [14] P. Li, M. Abdel-Aty, Q. Cai, and Z. Islam, ‘‘A deep learning approach 30 journal articles. His main research interests 697

639 to detect real-time vehicle maneuvers based on smartphone sensors,’’ include intelligent technology and application, 698

640 IEEE Trans. Intell. Transport. Syst., vol. 23, no. 4, pp. 3148–3157, numerical simulation, and fatigue analysis. 699
641 Oct. 2022.

98870 VOLUME 10, 2022

 S. Zhou et al.: Intelligent IoT Sensing System for Rail Vehicle Running States Based on TinyML

700 YONGKANG DU was born in Heze, Shandong, YONGHUA LI received the B.S. and M.S. degrees 719
701 China, in 2000. He received the bachelor’s degree from Liaoning Technical University, in 1995 and 720
702 in vehicle engineering from Weifang University, 2002, respectively, and the Ph.D. degree from the 721
703 in 2016, and the master’s degree in mechan- Dalian University of Technology, in 2005. She is 722
704 ics from Dalian Jiaotong University, in 2022. currently a Professor with the School of Locomo- 723
705 His research interests include machine learning, tive and Rolling Stock Engineering, Dalian Jiao- 724
706 embedded device development, and applications tong University, Dalian, China. She has published 725
707 of the IoT technologies. more than 50 journal articles. Her main research 726
interests include digital simulation, optimization 727
design of mechanical products, and fatigue relia- 728
bility analysis of vehicle structures. 729

708 BINGZHI CHEN received the B.S. and M.S.

709 degrees from Dalian Jiaotong University, Dalian, XINGSEN LUAN received the B.S. degree in 730
710 China, in 1994 and 1997, respectively, and the transportation from Ludong University, Yantai, 731
711 Ph.D. degree from the Dalian University of Tech- China, in 2021. He is currently pursuing the M.S. 732
712 nology, in 2002. He is currently a Professor with degree in vehicle engineering with Dalian Jiaotong 733
713 the School of Locomotive and Rolling Stock Engi- University, Dalian, China. His research interests 734
714 neering, Dalian Jiaotong University. He has pub- include machine learning and embedded device 735
715 lished more than 50 journal articles. His main development. 736
716 research interests include intelligent computing, 737
717 railway vehicle engineering, virtual prototype,
718 numerical simulation, structure optimization, and biomechanics.

VOLUME 10, 2022 98871